ORNVTM-13152 Design of Spatial Experiments: Model Fitting and Prediction Valerii V. Fedorov
ORNVTM-13152
Design of Spatial Experiments: Model Fitting and Prediction
Valerii V. Fedorov
This report has been reproduced directly from the best available copy.
Available to DOE and DOE contractors from the Office of Scientific and Techni- cal Information, P.O. Box 62, Oak Ridge, TN 37831; prices available from (615) 578-840 1, FTS 826-840 1.
Available to the public from the National Technical Information Service. U.S. Department of Commerce, 5285 Port Royal Rd.. Springfield, VA 22161.
This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, com- pleteness. or usefulness of any information, apparatus, product, or process dis- closed, or represents that Its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, does not necessarily conati- tute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.
ORNL/TM-13152
Computer Science and Mathematics Division
Mathematical Sciences Section
DESIGN OF SPATIAL EXPERIMENTS: MODEL FITTING AND
PREDICTION
Valerii V. Fedorov
Mathematical Sciences Section Computer Science and Mathematics Division Oak Ridge National Laboratory P. 0. Box 2008 Oak Ridge, Tennessee 37831-6367
Date Published: March 1996
Research sponsored by the Laboratory Directed Re- search and Development Program of Oak Ridge National Laboratory
Prepared by the Oak Ridge National Laboratory
Oak Ridge, Tennessee 37831 managed by
Lockheed Martin Energy Research Corp. for the
U.S. DEPARTMENT OF ENERGY under Contract No. DE-ACO5-96OR22464
.
C
DISCLAIMER
Portions of this document may be illegible in electronic image products. Images are produced from the best available original document.
.
DESIGN OF SPATIAL EXPERIMENTS ... 111
Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv Standard design problem . . . . . . . . . . . . . . . . . . . . . . . . . . v Optimal designs with bounded density . . . . . . . . . . . . . . . . . . x Correlated observational errors . . . . . . . . . . . . . . . . . . . . . . . xiv Random coefficients regression models: Trend estimation . . . . . . . . xviii Random coefficient regression models: Prediction . . . . . . . . . . . . xxv Comparison with the methods based on the variance - covariance struc- ture of observed random fields . . . . . . . . . . . . . . . . . . . . . . . xxix
xl Unknown covariance function . . . . . . . . . . . . . . . . . . . . . . . xliii Discrete case. Optimality criteria and the lower bounds . . . . . . . . .
10 Space and time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xlviii 11 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Abstract
The main objective of the paper is to describe and develop model ori-
ented methods and algorithms for the design of spatial experiments. Unlike
many other publications in this area, the approach proposed here is essen-
tially based on the ideas of convex design theory.
1. Introduction
Since the earliest days of the experimental design theory, a number of concepts like
split plots, strips, blocks, Latin squares, etc. (see Fisher (1947)), were strongly
related to experiments with spatially distributed or allocated treatments and ob-
servations. In this survey we confine ourself to what can be considered as an
intersection of ideas developed in the areas of response surface design of experi-
ments and spatial statistics.
The results which we are going to consider are also related to the results de-
veloped by Cambanis (1985), Cambanis and Su (1993), Matern (1986), Micchelli
and Wahba (1981), Sacks and Ylvisaker (1966, 1968, 1970) and Ylvisaker (1975,
1v
1987). What differs in the approach of this paper from those cited? We intend
to use the techniques which are based on the concept of regression models while
the cited studies are based on the ideas developed in the theory of stochastic
processes and the theory of integral approximation.
If this survey were to be written for a very applied audience, the title “optimal
allocations of sensors” or “optimal allocation of observing stations” could be
more appropriate. Environmental monitoring, meteorology, surveillance, some
industrial experiments and seismology are the most typical areas in which the
considered results may be applied. What are the most common features of the
experiments to be discussed?
1. There are variables x E X c Rk, which can be controlled. Usually k = 2 , and in the observing station problem, x1 and 22 are coordinates of stations
and X is a region where those stations may be allocated.
2. There exists a model describing the observed response(s) or dependent vari-
able(s) y. More specifically y and x are linked together by a model, which
may contain some stochastic components.
3. An experimenter or a practitioner can formulate the quantitative objective
function.
4. Once a station or a sensor is allocated a response y can be observed either
continuously or according to any given time schedule without any additional
significant expense.
5. Observations made at different sites may be correlated.
Assumptions 1 - 5 are very loosely formulated and they will be justified when
needed. In the subsequent sections the term ‘‘sensor” stands for what could be
an observing station, meteorological station, radiosonde or well in the particular
applied problem.
DESIGN OF SPATIAL EXPERIMENTS V
2. Standard design problem
In what follows we will mostly refer to experiments which are typical in environ-
mental monitoring setting as a background for the exposition of the main results.
We hope that the reader will be able to apply the ideas and techniques to other
types of experiments.
When assumptions 4 and 5 are not considered we have what will be be
called, the “standard design problem”. The problem was extensively discussed
(see for instance, Atkinson and Donev (1992), Fedorov (1972), Pazman (1986),
Pukelsheim (1994) and Silvey (1980)), and it is difficult to add anything new in
this area of experimental design theory. Theorem 1 which follows, is a gener-
alized version of the Kiefer-Wolfowitz equivalence theorem, (see Kiefer (1959))
and stated here for the reader’s convenience. It also serves as an opportunity
to introduce the notation, which is sometimes different from that used in other
articles of this volume.
Let
where 8 E R” are unknown parameters, fT( z) = (fi(z) . . . , fm(z)) are given
functions, supporting points x; are chosen from some set X , and the Eij are
uncorrelated random errors with zero means and variances equal to one. We
do not make distinctions in notation for random variables and their realizations
when it is not confusing.
For the best linear unbiased estimator of unknown parameters the accumu-
lated “precision” is described by the information matrix:
which is completely defined by the design [ = (.;,pi}?. In the context of the
vi
standard design theory 4
where ( (dz) is a probability measure with the supporting set belonging to X :
suppt c X, and
is the information matrix of an observation made at point x.
Regression model (2.1) and the subsequent comments do satisfy assumptions
1 and 2 from the previous section. To be consistent with assumption 3 let us
introduce a function 9 ( M ) , which is called the “criterion of optimality” in ex-
perimental design literature. A design
is called (@-) optimal.
Minimization must be over the set of all possible probability measures 2 with
supporting sets belonging to X . Now let us assume that:
(a) X is compact;
(b) f(z) are continuous functions in X , f E R”;
(c) Q ( M ) is a convex function and Q ( M ) 5 @(M + A), M 2 0, A 2 0,
Le. matrices M and A are nonnegative definite.
(d) there exists a real number q such that
(e) for any < E Z(Q) and f Z(q):
DESIGN OF SPATIAL EXPERIMENTS vii
where ~(a,<,f) = CY). Here and in what follows we use Q(<) for Q[M([)], Q" for @(<*) and min,,miq, J,
and so on, instead of minzrx, minteE7 Jx, respectively, if it does not lead to am-
biguity.
Theorem 1. . I f (a)-(e) hold, then
1. For any optimal design there exists a design with the same information
matrix which contains no more than n = m(m + 1)/2 supporting points.
2. A necessary and sufficient condition for a design <* to be optimal is fulfill-
ment of the inequality:
miny(x,f*) 2 0. 1:
3. The set of optimal designs is convex.
4. $(x, e") achieves zero almost everywhere in suppt',
where suppt, stands for supporting set of the design (measure) <.
Here and
Functions +(z, <) for the most popular criteria of optimality may be found, for
instance, in Atkinson and Fedorov (1984). Theorem 1 provides a starting point for
analytical exercises with various relatively simple regression problems and makes
possible the development of a number of simple numerical procedures for the
optimal design construction in more complicated and more realistic situations.
Most of these procedures are based on the following iterative scheme:
0 (a) There is a design & E Z(q). Find
x+ = argmin+(z,<,), x- = argmax$(z,Js), XEX XEX,
where X , = suppJs.
... VI11
0 (b) Choose 0 < ps < 1 and construct
[s+l = (1 - P S ) [ , + PS[(XS), where [(s,) is a measure atomized at 2,.
The choice of a sequence ,Os defines a variety of the algorithms; specific examples
are given by Atkinson and Donev (1992), Cook and Nachtsheim (1989), Fedorov
(1972, 1975) and Silvey (1980). The following sequences are most popular:
Theorem 1 together with iterative procedure (2.5), (2.6) provides quite power-
ful tools for constructing optimal design. The existing software products, see, for
instance, Mitchell (1974), Nguen and Miller (1992), Nachtsheim (1987), SAS/QC
Software (1995), Wheeler (1994), confirm this statement. Unfortunately, there
are a few hurdles, which do not allow the direct use of the results reported above.
The first one is that optimal designs defined by (2.4) may have unequal weights.
What does this mean in the context of observing stations allocation? If we have
N available stations or sensors, then r;* = [p:N] stations must be allocated at
s:, where b:lv] is some reasonable integer approximation of pzN. It is obvious
that in many cases (but not always) two or more stations sited in the immediate
vicinity of each other will not give essentially more information than a single
station. There are some arguments in favor of this statement, which can be
expressed economically in colloquial statistical terminology: observations from
these stations are strongly correlated. However, frequently weights pz may be
.
L
DESIGN OF SPATIAL EXPERIMENTS ix
considered as the desirable precision of measurements taken at the i-th station.
The corresponding precision can be achieved through the proper technical steps
or through controlling the longevity of the observational process.
Probably, Gribik et a1 (1976), were the first to use the optimal experimen-
tal design methods for environmental monitoring. They analyzed the problem
of allocating measuring resources to aid in accurately estimating ground level
pollution concentrations throughout a region X . The regression model was the
linearized version of the diffusion model for four pollution sources and unknown
background source. Since the diffusion model used in the study was a large scale
model, measurements separated by distances smaller than a threshold value dis-
tance appeared to be correlated in the corresponding parameter estimation prob-
lem. At the same time the design method was a particular case of the method
discussed in this section, where the independence of observational errors is essen-
tial. To avoid a contradiction the authors imposed the additional constraint: the
distance between any two observing sites must be greater than the characteristic
distance:
(Xi - "j)T("j - Zj) > d2.
Imposing constraints of that type is one of the simplest way to handle possible
correlation between the observed values at neighboring stations. Obviously the
approach does not work for long-range correlation, when the widely separated
observations are correlated.
It was assumed that the ground level pollution is of the prime interest. The
authors proposed to use the weighted average variance of the best linear unbiased
estimator of the ground level pollution:
as the criterion of optimality. The weight function W(Z) was selected proportional
to the population density in the considered region.
A rather detailed discussion of applicability of the standard design technique
x
for spatial experiments may be found in Fedorov et al( 1988).
3. Optimal designs with bounded density
Gribik et a1 (1976), used a very simple and transparent idea to avoid clustering
of sensors at particular points. This idea can be exploited in a more general and
formal setting. Let the number of sensors N be sufficiently large and the density
of stations per square unit be introduced into consideration:
Introduction of (3.1) is very reasonable when the sensor allocation is considered
in technological experiments. In the network allocation problem it is probably
less realistic. Nevertheless, the results considered in this section help to explain
why some intuitive approaches, similar to what was done by Gribik et a1 (1976),
do work well in most cases.
If X is not uniform (as might be appropriate say, with different topography
for different parts of X ) , then it is natural to assume that the sensor density has
to be constrained:
With obvious redefining of the design measure t ( d z ) and the upper bound @2(dz)
the latter may be reduced to a simpler statement:
Thus, the following optimization problem must be considered, (we skip the evi-
dent left hand side constraint):
(* = arg min 8 [ M ( [ ) ] , E
L
DESIGN OF SPATIAL EXPERIMENTS xi
This optimization problem was discussed by Wynn (1982), and Fedorov (1989).
To avoid unnecessary technical complications let us assume additionally to (a)-
(e) from section 2 that
(f) <P(dz) is atomless, Le.
The following theorem summarizes the most important properties of designs
with bounded density.
Theorem 2. . Let ZO be a set of design
((dx) > 0, and ((dx) = 0 otherwise, and let assumptions (a) - (f) hold. Then:
such that t ( d z ) = @(dx), when
- 0 There exists an optimal designs t* E LO.
0 A necessary and sufficient condition for this design to be optimal is that
$(x, [*) separates the two sets X * = supp[* and its complement.
In the above formulation "separate" means that there is a constant C such
that $ ( x , t * ) 5 C on X * and $(z,[+) > C on its complement.
Theorem 1 tells us that supporting sets of optimal designs must coincide with
the points where $(x,[*) achieves its minimum. Therefore, in most cases the
supporting set for the standard optimal design consists of a finite number of
supporting points.
Theorem 2 forces suppc to occupy the subsets of X . How is [*(dx) to be
realized by a practitioner? One of the possibilities is to replace t*(AX) for
relatively small areas AX by N * ( A X ) = [ [* (AX)N] . When N * ( A X ) is defined
then the corresponding number of sensors have to be allocated in A X . For
instance, they can be sited at the nodes of some uniform grid. Generally, that
allocation has to guarantee a reasonable approximation of the integral
xii
by the sum
The properties described by Theorem 2 allow us to formulate a simple numer-
ical algorithm to construct optimal designs (see Fedorov (1989)). Let a(&) =
$(z)dz and
lim a, = 0, lim a,! = 00 and lim < CQ. S+oO S+CO S+oO
s’=l s’=l
(a) There is a design [, E Eo. Let XI, = suppf, and X,, = X \ XIS. Two sets
D, C XI , and E, c X2, with equal measure,
and, correspondingly, including the points
(b) The design ts+l with the supporting set
is constructed.
Usually $(z) is assumed to be constant. All other cases may be converted to
this one with the proper coordinate transformation. In the computerized version
of the algorithm integrals in (a) are replaced with sums over some grid elements.
If these elements and subsequently Q, are fixed and elements of both D, and E,
coincide with the grid elements, then (a), (b) becomes an exchange type algorithm
(see, for instance, Mitchell (1974)) with the simple constraint: every grid element
cannot contain more than one supporting point and the weights of all supporting
points are the same, Le. N-l. In practice it is sometimes convenient to consider
... DESIGN OF SPATIAL EXPERIMENTS Xl l l
grids of varying density, which has to be proportional to #(z). While it can be
shown that the exchange algorithm (a), (b) converges to an optimal design for
properly diminishing a, , it is not generally true for finite a, and, in particular,
when as N - l . The accuracy of the limit designs is defined by the accuracy of
the approximation (see assumption (e) from Section 2):
r
When these approximations are reliable enough then we can hope that the
limit designs do not deviate too much from the optimal ones. The term “limit
design” must be used with some reservation when a, E N-’: instead of conver-
gence some minor oscillations of @ [M(t , )] may be observed. Practical aspects of
the iterative procedure (a),(b) were discussed by Fedorov and Muller( 198913) in
the air pollution network design setting.
4. Correlated observational errors
Let us assume now that the random errors in model (2.1) are correlated and
that the covariance structure is known, i.e. either the covariance matrix V or the
covariance function V ( z, z’) is given. There is no need to use the second subscript
indicating the repeated observations and we consider
where i = 1 , . . . , N , E(&;) 0 and
For the obvious reason, in this section we will use the simplified notation:
xiv
In the case of correlated observations the best linear unbiased estimator is defined
as ( see, for instance, Rao (1973)):
where
where
Unlike (2.2), the information matrix (4.4) is not a sum of information matrices
of single observations. Therefore we cannot use directly the results of the convex
design theory, which is essentially based on the additivity of information matrices.
DESIGN OF SPATIAL EXPERIMENTS
Actually, we have to consider the optimization problem
[G = arg min @ [M(e)J , (4.7) €N
which does not have too much in common with (5) besides notation. For instance,
the convexity of @ is not very helpful anymore.
In most studies authors try to imitate the iterative methods of optimal design
construction considered in two previous sections. For instance, computations
become similar to the standard (uncorrelated) case, if the following recursion
formula is used (Brimkulov et a1 (1986)):
We can easily derive, for instance, that
Subsequently, for the D-criterion the point
must be added to the design &,T. That is an imitation of step (a) from the iterative
procedures considered in the two previous sections.
There exists a simple intuitive explanation why iterative procedures based on
(4.10) provide "good" supporting points in the sense of the D-criterion. First,
let us recollect that in the no-correlation case accordingly to stage (a) of the
iterative procedure from Section 2 the additional observation( s) must be allocated
xvi
at point(s), where the ratio
variance of prediction with the estimated 8 variance of prediction with the given 8
- - a2(x> + fT(z>N-- l (5) f (4 a2(4
is maximal. This follows, for instance, from (2.5) when
(4.11)
in the more general case (see Fedorov (1972)) for details. In the case of correlated
observations we are looking for a maximum of the same ratio
variance of prediction with the estimated 8 variance of prediction with the given 8
When z + xi E s u p p f ~ , then
for f(z) and V ( x , JN) continuous in the vicinity of 2. In other words the iterative
procedure defined by (4.10) does not admit coinciding supporting points. The
result follows from the definitions of +(z,[~) and S2(z,[~), and the fact that
where S;j is the Kronecker symbol. Formula (4.10) can be easily rewritten for
the deleting procedure. LFrom (4.9) it follows that in the case of the D-criterion
(4.12)
DESIGN OF SPATIAL EXPERIMENTS xvii
candidates for deleting are defined by the equation:
We are not aware of any results on the properties of the iterative procedures
based on (4.10) and (4.14) for the D-criterion or similar procedures for other cri-
teria. There are empirical confirmations that the exchange-type algorithms lead
to a significant improvement of the starting design. For instance? Rabinowitz and
Steinberg (1990) applied that type of algorithm to the problem of selecting sites
for a seismographic network. They have shown that the computed designs are rel-
atively efficient and are better than the standard D-optimal designs constructed
for models with uncorrelated observations. It is reasonable to note that compu-
tationally (4.9) and (4.14) are much more demanding than their counterparts in
the standard design theory. There exist a number of studies where the optimiza-
tion problem (4.1) is considered for some special and relatively simple covariance
functions? for instance, generated by autoregressive models. Various details and
further references may be found in Bickel and Herzberg (1979), Bishoff (1992),
Kunert (1988), Martin (1986), Miller and PBzman (1995) and Niither (1985).
In conclusion of this section let us emphasize again the significant difference
between the case with correlated observations and the standard case. For un-
correlated observations the additiveness of the normalized information matrix
(a(x) = 1): IV
A!!([) = N-l f ( x ; ) f T ( x ; ) = N-lM(<N) i=l
leads to many simple and elegant theoretical results initiated by Kiefer’s pioneer-
ing findings. Very frequently normalized information matrices may be treated as
a limit, i.e.:
(4.15)
In many cases for correlated observations the corresponding limit does not exist
and the matrix M ( [ ) cannot be introduced. One of the most successful attempts
xviii
to replace (4.15) was due to by Sacks and Ylvisaker (1966, 1968); see more in
Section 7.
5. Random coefficients regression models: Trend estima-
t ion
In what follows we intend to consider some simple models for the random com-
ponent in (2,l). It is convenient to partition “intrinsic” or “proce~s’~, and “ob-
servational” sources of randomness:
Values u;j describes deviations of the observed response from q(x;, 0) due to
some causes which are independent of an observer. For instance, an average wind
velocity may be disturbed by various local micro-eddies. The term e;j describes
L‘observationa17’ errors. Sometimes these errors are defined by the selected obser-
vational technique and, at least partly, they are controlled by an observer. The
proposed partitioning is very conditional, and the reader may use a different one,
which is more compatible with the corresponding experimental situation.
Let us assume that
or
where
.
DESIGN OF SPATIAL EXPERIMENTS
Vector 0 2 is random with
E(&) = 0, E(620:) = Var(02) = A,
vector e describes the observational errors, which are random and
E ( € ) = 0 E(&) = 021.
We assume that 82 and E are uncorrelated. In terms of (5.1) we have
E ( u ) = 0, E ( U U T ) = F , T ( t N ) A F 2 ( t N ) .
xix
(5.4)
If e = u + E , then
Thus we are going to consider a very special case of (4.1) with V(&v) defined by
(5 .5) . It may be illuminating to associate index “j” with time (hour, day, . . .) and “i” with location (z; is a vector of coordinates of a particular site).
Model (5 .5 ) gives an opportunity to introduce criteria of optimality which
provide a very reasonable description of various experimental situations. Those
criteria may be divided in two main groups. The first group is related to the
“average over time” behavior of the observed response. The corresponding criteria
depend upon the precision of estimators of 01. This means that we consider some
functions of Var(&l), where 81 is an estimator of 61.
The second group deals with “instant” responses and the corresponding cri-
teria are based on V a ~ ( e ) , eT = (e:, 6;).
Let us start with the first group, i.e. with estimating the subvector 61. The best linear unbiased estimator is (compare with (4.3):
xx
The dispersion matrix of 4 is
where
An optimal design (or optimal observational network) is defined as
(5.7)
which differs from (4.7) only by the more detailed information about V ( ~ N ) .
It may be expedient to note that unlike the situation described in comments
accompanying (4.13) the covariance is not anymore a continuous function at the
diagonal:
x+x/ lim E(e(z)e(z ' ) ) # o2 + f . ( ~ ) A f i ( z ' ) . (5.10)
Therefore (5.9) may admit designs with repeated observations, i.e. it could
be that x E supp&v and M ( & + 2) is better than M(JN) .
Using the identity
(5.11)
(5.12)
DESIGN OF SPATIAL EXPERIMENTS xxi
LFrorn the Frobenius formula it follows that
The matrix D ( ( N ) can be also considered as the dispersion matrix of the best
linear unbiased estimator of parameter OT = (e:, O r ) for the regression model.
where F T ( h ) = ( F : ( I ~ N ) , F . ( ( N ) ) , E ( € ) = O,E(ceT) = o21, with the prior
information about parameters 02 described by a prior distribution P(82) such
that
and
See, for instance, Fedorov (1972)? Pilz (1991), and Seber (1977). Thus, the
optimization problem (4.1) may be embedded in the framework of convex design
theory. For instance, for the D-criterion, when (Ml1(&v)I-l = IDl1(&)( must be
minimized, one can use any algorithm developed for the construction of “exact”
or “discrete” optimal designs; see, for instance, Cook and Nachtsheim (1980),
Fedorov (1972), Ermakov (1983), and Pukelsheim (1993) when only the subvector
81 has to be estimated. More generally we can now describe experimental design
as the following optimization problem
(5.13)
where
N is now the total number of possible observations, and
xxii
Let us note that occasionally the total number of observations and the number
of supporting points may coincide (like in (5.2). Then N stands for both. The
results of Sections 2 and 3 may routinely be applied to (5.13) when 8 is properly
defined and o2 and A are known.
Subset D-optimality. According to (4.10) we have to minimize some function
of the matrix a;:, when the parameters 01, are of prime interest. In terms of
(5.13) it means that the objective function 8 must depend upon elements of the
matrix All([), which may be defined as follows:
where
iwOz2 = a2 N - ~ A - ~ .
One of the possibilities is to select
(5.14)
LFrorn Theorem 1 it immediately follows that a necessary and sufficient con-
dition for [* to be optimal is that (compare with Fedorov, 1972):
(5.15)
Notice that
DESIGN OF SPATIAL EXPERIMENTS xxiii
is the normalized variance of JTf(s), where 8 is the best linear unbiased estimator
of 8 from model (5.2), and
may be considered as a normalized variance of the best linear estimator for
the regression model with the same observational errors but with the response
0; f2(x). To get non-normalized values we have to multiply the normalized values
by a2N- l . Simple, but rather long matrix calculations show that
(5.16)
(5.17)
This collection of formulae looks much more complicated than the similar terms
in (5.15). However, that presentation has one remarkable feature: it does not
depend upon functions f2(z) and matrix A explicitly. All elements in (5.16) are
xxiv
completely defined by the covariance function
o - ~ N R ( z , 5') = E (f2(~)6:62f, T I (Z )) = V(Z,Z'), # 5'-
(5.18)
Thus, when the covariance function is known directly, i.e. we do not use (5.5)
to get it, one can use (5.16) and (5.18) to construct optimal design. Moreover,
the cases, in which E = dim62 --+ co may be considered. The first attempt in this
direction was done by Miiller-Gronbach (1993).
r
6. Random coefficient regression models: Prediction
The presentation of the design problem for model (5.2)-(5.4) in the form (5.13)
allows us to develop a rather simple technique for experimental design when the
objective is the prediction of observed values. For the sake of simplicity, let
~ ( z , O ) F 0 in (5.2) and
~ j ( ~ i ) = uij = 6 r f ( ~ i ) .
Then the corresponding optimal designs are defined as
[* = argminQ ( M ( [ ) + Mo) , 5
where
M ( [ ) = / f(z)fT(z)[(dz) and Mo = a2N-lA- l .
It is expedient to note that
DESIGN OF SPATIAL EXPERIMENTS XXV
where the expectation operator E takes into account randomness of both obser-
vational errors and regression parameters. Minimization is taken with respect
to all linear estimators, see Gladitz and Pilz (1982), Fedorov and Miiller (1989),
Pilz (1991) . The best linear estimator is
Similar to arguments in Section 5 we can apply the equivalence theorem to
(6.2) to find, for instance, necessary and sufficient conditions for a design [* to be
optimal. Leaving to the reader the possibility to formulate them for the general
case we focus only on three simple and very popular criteria.
Minimax and D-criterion. For the D-criterion, when
one can easily derive from Theorem 1 that a necessary and sufficient condition
for to be optimal is that
for all x E X .
This inequality appears, especially when the dimension of f is large, more at-
tractive and meaningful in notation described in comments accompanying (5.11):
The variance of the best linear unbiased predictor for u(x) equals
xxvi
= c2 + Cr2N-l [R(z,z) - RT(z7E) (w-1 + R([))-l R(z7E)] ,
where
G(X) = OTf(z) = Rt(2,J) (w-1 + R([))-lP-
= VT(", [) (a2W-1N4 + v(())-I F,
and the components of the vector P are averages of observations at the corre-
sponding points. For the continuously changing weights pi the i-th component of
Y may be considered as the observation made with a precision a/p;N = ai2. If
one introduce the covariance function
-
then (compare with (4.6) or coming later (7.1)) predictor &(z) coincides with
the best linear unbiased predictor for y(z) = u(z) + E(X) everywhere except
x E suppt. At the design points xi the realization(s) of y(z) are measured
directly and are not needed to be predicted, i.e. one may select y(z) = ij(x) and
Var (y(x) - i j(z)I() = 0. Obviously
(6.10)
otherwise.
Using Theorem 1 together with (6.5) and (6.10) we can formulate the analogue
of the Kiefer-Wolfowitz equivalence theorem:
Theorem 3. The following two design problems are equivalent:
There is one significant difference between this result and the original equivalence
DESIGN OF SPATIAL EXPERIMENTS xxvii
theorem: an optimal design generally depends upon the number of observations
N to be used.
Theorem 3 and formula (6.5) give another insight into numerical procedures
from Sections 2 and 3: at every stage one has to relocate the design measure from
the point(s) where y(z) may be predicted easily (small Var(y ( z ) - $(z)[t,)) to
the point(s), where the prediction is poor ( large Var(y(z) - $(x)[&)) . Two linear criteria. Two objective functions which are very popular in spatial
statistics are the weighted average variance of prediction:
and the variance of the weighted average of prediction:
where 2 is the “prediction” set or the area of interest. Using (6.6) one can find
that minimization of Q1(t) and Q2(() is equivalent to minimization of
where in the first case
and in the second one
A = J, w2(z)f(z)fT(z)dz,
A = aaT, a = w(z)fT(x)dz.
LFrom part 2 of Theorem 1 it is easy to conclude that
Theorem 4. . The design e* is linear optimal if and only if
(6.11)
(6.12)
Similar to the D-criterion we can show that for the average variance of pre-
diction
4(z7 5) = 41(3>[) = 1 c0v2(z, Z ' l < ) W 2 ( z 1 ) d Z r , (6.13) z
while for the variance of the weighted average
where
C0v(z7d1t) = R ( z , d ) - RT(.,t) (w-1+ R(J))-I R ( d 7 t ) .
The counterparts of Theorem 3 and 4 may be formulated for optimal designs
with bounded density. To do this function $(z,[*) in Theorem 2 should be
replaced either with Var (y(z) - $(z)It), or with ~I(z, t) , or with +2(z7 5). Remarks on applicability of the results. Let us note that the introduction of
model (5.2) -(5.4) to generate correlated observations allows us to use the convex
design theory for regression problems with correlated observations. Moreover, all
results may be presented in a form which does not demand any direct knowledge
of the functions f(z). We can formulate results for a particular criterion using
only information about the covariance function.
In this and in the previous section we have discussed only the properties of
optimal designs. We hope, that having the sensitivity function $(x, [) represented
for various criteria in terms of the normalized covariance function Cov(z, z ' I ~ ) , the reader can easily construct numerical procedures similar to those discussed
in sections 2 and 3.
7. Comparison with the methods based on the variance - covariance structure of observed random fields
Sacks-Y"visarEer approach. Let us suppose that in model (4.1) there is no trend,
DESIGN OF SPATIAL EXPERIMENTS xxix
i.e. v ( x 7 8 ) E 0, and the covariance function V(x ,x ’ ) is defined and known for
all x,x’ E X . The objective of an experiment is to predict y(x) at a given set of
points 2, which can be either discrete or continuous.,
The best linear unbiased predictor for y(x) may be presented as follows (com-
pare with (4.5 and (4.6)):
We again use notation JN = (x l , . . . ZN) to emphasize that there is only one
observation at every point 2;. Criteria Q 1 ( ( ~ ) and Q2(&) introduced in the
previous section have been most intensively analyzed in the studies related to the
design problem with correlated errors. Usually it has been assumed that 2 = X .
A very good summary of the main results for the criterion Q1((~) may be
found in Micchelli and Wahba (1981). The criterion Q 2 ( & ) was analyzed by
Sacks and Ylvisaker (1970) and Ylvisaker (1987). Further references and com-
ments may be found in Cambanis and Benhenni (1992), Cambanis (1985), Cam-
banis and Su (1993).
Noting (see (7.l)that
where
we can consider minimization of either Q1(&) or & 2 ( [ ~ ) as a problem of finding
an optimal basis for a quadrature formula in approximation theory, with a rather
specific objective function; see Karlin and Studden (1966), Sacks and Ylvisaker
(1970), Stroud (1975).
When N ---f 00, both &I(&) and Q 2 ( [ ~ ) converge to zero for “smooth”
xxx
covariance functions V(z,x') and for any atomless sequence JN. As in Section
3 we may introduce the limit design measure that defines JN. How a sequence
(N may be generated with a particular [ (dz ) is discussed in details by Cambanis
(1985) for X C R1. For instance, the so called regular median sequence or design
.& is defined as
[(dx) = 2N , 2 = l , 2 , . . . , n ,X = [a, b]. (7.3) . X N ; = arg
When X is hypercube and V(z ,x ' ) is separable with respect to all components
of 2, then design [ N may be defined as a direct product of univariate designs (see
Ylvisaker (1975) for details). Thus the design problem is reduced to the search
of the limiting measures providing the best convergence rate for the selected op-
timality criterion. The rather elaborate technique, a close sibling of the classical
approximation theory, leads to a very special minimization problem. Introducing
the design density t ( dz ) = h(s)dx we may state this problem as follows
h* = arg min Q [B(h)] , h (7.4)
where
m is the number of estimated integrals (for instance, integrals of Q2((~)-type
with various weight functions), functions ~,(z) and integer k are defined by a
covariance function and by an optimality criterion.
Similar problems were considered in studies concerned with simultaneous cal-
culation of m integrals by Monte-Carlo method; see Mikhailov and Zhigljavsky
(1989), Zhigljavsky (1988), for details and further references. Actually, (7.4) be-
ing an optimization problem in a space of probability measures has many features
in common with the standard design problem. Some interesting results including
the analogue of the iterative numerical procedure from Section 2 are summarized
DESIGN OF SPATIAL EXPERIMENTS xxxi
and discussed by Zhigljavsky (1988).
. Analytical solutions of (7.4) for the one-dimension case were proposed in the
pioneering papers by Sacks and Ylvisaker (1966, 1968, 1970). Various general-
izations may be found in Hajek and Kimeldorf (1976) and Wahba (1971, 1974).
Following Cambanis (1985), the essence of those findings can be formulated as
follows:
If there exist exactly k quadratic mean derivatives of the random process y(z),
then under certain regularity conditions (see details in the cited publications)
where
and superscripts indicate the order of partial derivatives. For any design with
density separated from zero the integral Q(&,T) diminishes as O(N-2k-2) and
h* (z) minimizes
lim N2k+2 Q ~ ( < N ) ( 7 4 N-bcr,
In fact, the immediateobjective of Sacks and Ylvisaker (1966,1968,1970) was min-
imization of some function of the dispersion matrix of estimators of parameters 6
describing a linear trend 8’f(z) in model (4.1). They reduced the corresponding
minimization problem to minimization of objective functions similar to Q2( &v).
For instance, when 8 f R1, then one has to minimize Q ~ ( [ N ) with a weight
function which is a solution of the following integral equation:
f(z) = J , V(z, z’)w(z’)dz’.
Evidently, for stationary covariance functions ak(z) constant. Subsequently
the optimal limiting density h*(x) is completely defined by the weight function
w(z). In other words, only the behavior of V(z,z’) at its diagonal influences the
solution! The Sacks-Ylvisaker approach (at least in its current form) cannot be
xxxii
used in two cases of practical importance. First, it does not work for “infinitely”
smooth covariance functions, when ak(z) 0 for any IC. The covariance function
is a popular example (see Sacks and Ylvisaker (1966). The presence of the “white”
noise (see model (5.1)) in observed variables gives another example, when the
approach does not work. The latter case is of interest for many applications
being a very reasonable model when a random process is observed with some
instrumental error. In conclusion of this subsection let us note that the concept
asymptotically optimal design [$ based on the existence of a continuous limit
density h*(x) and assumption that
lim max(z;N - = 0;
see Sacks and Ylvisaker (1966). The definition (7.3) of design ,& is one-dimension
by its nature and that makes the approach difficult for spatial applications; see
Ylvisaker (1975) for further details.
N-CQ a
Random parameters approach, To understand the advantages and disadvan-
tages of approach proposed in Sections 5 and 6 relative to to the Sacks-Ylvisaker
approach let us introduce the following model:
where fa(z) are eigenfunctions of the covariance kernel
6, are random with zero means and diagonal covariance matrix, such that Var(6,) =
A,, A1 2 A2 L . . . 2 A, 2 . . . , e(z) is white noise with the variance 1, and o2
is a normalizing constant.
It is well known (Mercer’s theorem; see for instance, Kanwal (1971)) that
DESIGN OF SPATIAL EXPERIMENTS xxxiii
under very mild conditions the series
is uniformly and absolutely convergent and subsequently under very mild as-
sumptions {A,} must diminish not slower then O(l/cr) . For many widely used
stochastic processes or fields the rate is significantly faster; see Micchelli and
Wahba (1989, Theorem 3.
Therefore, for sufficiently large n the kernel Vn(x, 2’) may be a very reasonable
approximation of V ( x , x‘). Allowing 0 +. 0 we may hope that the process yn,@(x)
is “close” to y(x) in the sense of their second moments. Subsequently, we might
expect the closeness of the corresponding optimal designs. This probably holds for
designs 5~ with relatively small N . However, for N + 03 the diminishing Q does
not guarantee closeness of optimal designs with 0 = 0 and 0 > 0. First, formally
the Sacks-Ylvisaker approach does not work for any model with additive “white”
noise, because it causes discontinuity of a covariance kernel at its diagonal:
Secondly, for any cr > 0 and any [ N the rate of convergence for either Q ~ ( [ N ) or
for & ( [ N ) will not be generally better than O(N-’) . This is slower than for any
continuous covariance kernel.
Thus, for large N the Sacks-Ylvisaker approach and results from Sections
5 and 6 may lead to the different asymptotically optimal designs. If one be-
lieves that there is no instrument or any other observation error, then the Sacks-
Ylvisaker approach leads to the better limit designs.
When the contribution of observation errors is significant then approximation
(7.8) becomes very realistic and allows the use of methods from Sections 5 and 6,
which usually produce optimal designs with very moderate numbers of support-
ing points. Usually these designs have about n supporting points.The existence
xxxiv
of well developed numerical procedures and software allows the construction of
optimal designs for any reasonable covariance function V(z7 5’) and various de-
sign regions X, including two and three dimension cases. Let us notice that the
function COV(Z, 2’15) used in Theorem 4 may be presented in the following form:
(7.10)
P = 6;jri7 r; = p;N.
This formula is convenient for some theoretical exercises. For more applied ob-
jectives and for development of numerical algorithms based on the iterative pro-
cedures from Sections 2 and 3, the direct use of eigenfunctions f & ( ~ ) is more
convenient.
Popular IcernekThere are several show-case processes and design regions for
which analytic expressions for the covariance kernel exist, and the corresponding
eigenvalues and eigenfunctions are known:
For the Brownian motion the kernel is
V(Z, z’) = min(z, z’), O 5 x,x‘ 5 1,
and its eigenvalues and eigenfunctions are
A, = (a - 1/2)-’~-’, f,(z) = &sin(a - 1/2)ns, Q = 1,2,. . . .
For the Brownian bridge,
and -2 -2 A, = a R , f,(s) = d i s i n a n z , a = 1,2 ,....
Both kernels are not differentiable on the diagonal (see comments to (7.5)) and
their “jump” functions Q(Z) are easy to calculate. These two kernels or some
DESIGN OF SPATIAL EXPERIMENTS xxxv
simple functionals of them (compare with Wahba (1971)) are convenient candi-
dates for the Sacks-Ylvisaker approach. For the Poisson kernel
1 - p 2
1 - 2pcos27r(J: - xc’) + p 2 ’ V ( X , d ) = 0 2 x,xI 5 1, 0 < p < 1,
and
A* = 1, x 2 , 4 = A2, = pa,
The shape of the Poisson kernel may be controlled by the parameter p. It is
%mooth” at the diagonal and the Sacks-Ylvisaker approach cannot be used.
For most real-world problems it is impossible to represent covariance kernels
in a simple closed form. However, a representation in the form of an infinite series
is standard. For instance, in many experiments related to either diffusion or heat
conduction the covariance kernel may be expressed in the the two dimensional
finite domain case (see Butkovskiy (1982)) as
where X = (0 5 q , x 2 5 l}. Evidently,
fap(s) = 2 sin azrI sin p7rx2, ~ , p = exp [-a2r2(a2 + p 2 ) ] , a, p = I, 2, . . . ,
and a is some constant. Representation (7.11) is very natural and convenient for
the techniques considered in Sections 5 and 6. Note that the physical problems
mentioned above lead to the Gaussian type kernels (compare with (7.7)) when
X becomes infinite with respect to any or both coordinates.
The curious reader will find more covariance kernels in any serious book on
integral equations containing a chapter on definite kernels or describing Green’s
functions (see, for instance, Kanwal (1971) and Butkovskiy (1982)).
Selecting the number of terms in (7.8) sufficiently large assures the closeness
xxxvi
of Vn(z, d) and V ( x , z') may be assured. In many cases it is convenient to assume
that coefficients {e,}? are normally distributed. This assumption does not help
in the present problem and in fact can cause some theoretical difficulties. To
avoid that we suppose that for any Q: distribution P(8,) has a finite support set
[a,, b,] in R1. For instance, we may select some simple symmetric distribution
P ( 0 ) defined on f-a,a] with E(02) = 1 and use P(B,d,) as a distribution for
6,. Selection of distributions with finite supporting sets assures not only close-
ness of an exact kernel and its approximation but proximity of C,"=18afa(z)
and c'& 6,f,(z) if the sequence {f,(s)}? satisfies some routine assumptions
from the approximation theory. The basic idea of using model (7.8) is in deriv-
ing optimal designs for the approximate model 8,f,(z) and verifying the
fact that these designs are optimal or close to optimal for Vn(z,d) or V ( z , d ) .
Furthermore, if the objective function is uniquely defined by a dispersion matrix
of estimated parameters, then the constructed design is optimal for any model
identical to the used one in terms of the first and second moments.
Using (7.8) with o = 0 we may immediately conclude that minimization of
&((AT) is a rather standard problem from the approximate integration theory,
see Davis and Rabinowitz (1985), Stroud (1975).
For instance, it is known (the Gauss-Jacobi Theorem) that for any polynomial
p(x) of degree k 5 2N - 1 the exact equality
.-h N
(7.12)
can be achieved the properly selected weights and supporting points. If { p , ( z ) } f
are orthogonal polynomials with the weight function w(z) > 0,then [AT = {zi)?
are zeros of p j ~ ( z), and
N-1 . -
!IT1 = 4;l(h) = E pi ( . ; ) , (7.13)
DESIGN OF SPATIAL EXPERIMENTS
For example, let us consider the Brownian bridge kernel. Note that
~~
xxxvii
(7.14)
where t = COSTX and Ua-l(t) is the second kind Tschebysheff polynomial. Now,
with z ( t ) = T-’ arccos t , we have for a proper <N = {xi)?
where y,(x) = E:=, OcYfa(x), and it follows from (7.14) that y, ( ~ ( t ) ) / d n is a polynomial of degree not higher than n - 1. If w ( ~ ( t ) ) = d m , then s;V must coincide with zeros of U N ( ~ ) , which are
i N + 1 ’
2; = .(ti) = - i = 1, ..., Nl 2N 2 n;
compare with Muller-Gronbach (1993). Accordingly to (7.13) weights are q ; ( < ~ ) =
T sin2 ~ s i / ( N + l ) , and finally
The solution is extremely simple, but it could be more complicated for other
weight functions, see Davis and Rabinowitz (1986). The value of Q2(&) is of or-
der 0( X2N). Various results about remainders in approximate integration theory
may lead to the better estimates, but the corresponding technique is beyond the
scope of this paper. Further details and related results may be found for instance,
for the one dimension case in Davis and Rabinowitz (1986), Szego (1959) and for
the multi-dimension case in Stroud (1975). Similar exercises may be done for
the criterion & 1 ( e ~ ) with Q = 0. The minimization of Q 1 ( t ~ ) now becomes a
problem from the function approximation theory. When 2 = X and w(z) z 1,
xxxviii
then it follows from (7.1) that
n r \ 2
where vT(x) = V(Z,[N)V-~(&). It is known (see e.g. Micchelli and Wahba (1981)), that
= 2 A,. a=N+1
(7.16)
This lower bound may be used to evaluate the efficiency of JN and can be achieved
for any singular kernel, i.e. when
To verify the latter conjecture one has to select the design [N coinciding with
all zeros of fX(z ) ; see some additional details in Fedorov and Hack1 (1994). In
cases when eigenfunctions cannot be found analytically the use of the remainder
theory is probably one of the most reliable ways to construct satisfactory designs;
see e.g. Davis and Rabinowitz (1984) or Achieser (1956). The ideas discussed
in this subsection help to generate effective designs with very moderate number
of observations N , obviously much less than we need using the Sacks-Ylvisaker
approach based on the local approximation of y(z). The author is not familiar
with any studies where the connection between the classical approximation theory
and the Sacks-Ylvisaker approach were analyzed systematically for models of type
(7.8) with n + 00. Perhaps Micchelli and Wahba (1981) and Miiller-Gronbach
(1993) considered the closest ideas and models.
Again, we would like to note that in most cases measurement errors may
contribute substantially to the randomness of observations. The rule of thumb in
selection of the number of terms in (7.8) is that the least eigenvalue AN should
be significantly less than 0.
xxxix DESIGN OF SPATIAL EXPERIMENTS
8. Discrete case. Optimality criteria and the lower bounds
When the design region X and the set of interest 2 are discrete and contain
Nx and Nz points correspondingly, then the covariance matrix of the vector
{y(x;) - fj(xi)}p, z E 2, completely describes any objective function based
on the second moments. We use the notation D z ( t ~ ) , when the latter matrix
consists of elements
The discrete versions of &I(&v) and & 2 ( t ~ ) are correspondingly
where LT = (1,. . . , 1). We will introduce any weights as we did in the continuous
case, to keep notations simple. In the discrete case we may introduce a very
special version of D-op t imality
It is assumed that there are no points in common for 2 and s u p p t ~ . Otherwise
the determinant equals zero, because
when x; E s u p p t ~ .
Criterion (8.2) is very popular in the statistical literature related to the opti-
mization of monitoring networks; see, for instance, Guttorp et a1 (1993), Carelton
et a1 (1992), Schumacher and Zidek (1993), Shewry and Wynn (1987). In the
cited papers the authors talk about either entropy or information. After the as-
sumption of multivariate normality of the corresponding distributions is made, all
approaches lead to various modifications of D-optimality; compare with Lindley
xl
(1956).
In addition to (8.1) and (8.2) a number of other criteria were introduced for
application in monitoring network improvement. A good collection of them can
be found in Megreditchan (1979, 1989); see also Fedorov and Hackl (1994).
As soon as the criterion of optimality &(<N) (we use this notation to emphasize
that only the criteria of optimality related to the problem of interpolation or
extrapolation are considered in this section) and the kernel V ( z , z’) are defined,
we have to find
E; = arg min &( h). (8.3) t N
It is interestingly to note that optimization problem (8.3) was considered
in a very different setting by Currin et a1 (1991) and by Morris et a1 (1993);
see also Sacks et a1 (1989) for older references. They considered the Bayesian
approach to design of computer experiments and introduced Q(<N) as a measure
of discrepancy between a computer model and its approximation based on some
prior knowledge expressed through the smoothness of the exact response. The
latter was defined by a covariance function.
When N z is relatively “small” and Nx is not very “large” then exhaustive
search may be a proper numerical procedure for a modern computer. With
increase of NZ and NX one can use the exchange type algorithms discussed in
Shewry and Wynn (1987) and in Fedorov and Hackl (1994), which are similar to
those discussed in sections 2-4.
An alternative approach may be based on the introduction of a model similar
to (7.8). For the sake of simplicity of notations, let 2 c X , and let
where XI 2 A2 2 . . . 2 X N ~ , Kj = V(zi,zj) and q , x j E X. Then one may
consider the following approximate model
N
Y 22 YN = e,f, (8-5)
DESIGN OF SPATIAL EXPERIMENTS Xl i
where all vectors Y, YN and fff have Nx components and
Similar to (7.16)
NX
cu=N+l = [(Y-yN)T(Y-yN)] = &Y, (8.6)
where F = (f1,. . . f N ) . There exists another result that can help to evaluate
the closeness of I'v and Y. Let
N
Then (compare with Rao (1973), Ch. 8g)
VN = argmin /IV - All, A
where rank A = N and symbol llBl/ denotes the Frobenius norm of B defined by
(trB2)'i2 = (Eij B;)li2 . Moreover
Thus, the vector YN is the best (maybe not unique) approximation of Y in
the sense of two criteria (8.6) and (8.7). In fact, it is the best one for any strictly
increasing function of D = E [(Y - ?N)(Y - YN)'] which is invariant under
orthogonal transformations, where $$ = BY, rank B 5 N ; see Seber (1984),
Ch. 5.2. The vector
FN = V ( X , [ N ) V - ' ( [ N ) Y 7
where V T ( X , t ~ ) = ( V ( q , 5 N ) ,... ,V(ZN,[N)) , is one of the above linear es-
xlii
timates. Therefore (8.6) and (8.7) help to find the lower bounds for criteria
depending upon
Model (8.5) helps to understand some features of optimal designs and lead to
some interesting numerical procedures (see next section). Adding the “white”
noise, i.e. introducing the following model
N
where E(&) = 0 and E(&‘) = I , allows us to use all the tools discussed in
Sections 2, 3, 5 to generate optimal designs.
9. Unknown covariance function
All the results discussed in the previous sections have essentially used the fact that
either a covariance function V ( z , z’) or a matrix A is known. That is possible but
unfortunately uncommon in practice. In this section we explore two approaches
to estimate the covariance structure.
Direct estimation of a covariance matrix. Let us start with a discrete design
region X and assume there exist repeated observations at every point of X .
Meteorological and environmental networks provide the most typical examples;
see e.g. Megreditchan (1979, 1989) and Oehlert( 1995a,b).
residuals Y - p as Let us define (compare with the previous section) the dispersion matrix of
Where J N = (21,. -. ,zN), y T ( h ) = (~(zI), . . . , y ( z ~ ) ) , and B is an Nx x N
matrix. For the sake of simplicity we assume that E ( Y ) = 0 and that this fact
DESIGN OF SPATIAL EXPERIMENTS xliii
is a priori known. Minimization is understood in the matrix ordering sense. A
solution of (9.1) is
B* = v(x,tN)v-l(tN)
When sufficiently many observations are accumulated at every point of X the
strong law of large numbers assures us that
(9-2) and subsequently B* and B*, which minimize correspondingly the left and right
hand-sides, are close to each other. Straightforward minimization gives
where both matrices with caps are evident partitions of
k
Q ( X ) = q y . j=1
(9-4)
When there are missing observations, then it is better to use instead of (9.4)
pairwise estimates
where
and
kil
j=1
& ( X ) = ICif'Cy,iT(,l,
IC;! is the number of cases when the response variable was measured at xi
simultaneously.
Thus, (9.2) - (9.4) lead us to a very simple and widely used recipe: replace
unknown parameters by their estimates and use methods developed for cases in
which all parameters are known. Together with this simple recommendation (9.2)
helps to generate other versions of numerical algorithms considered earlier. Let
xliv
us introduce matrix I ( J N ) with the following elements:
Sii, when x; f supp J N ,
0, otherwise. IiZ(<N) =
The left-hand side of (9.2) may be represented now as
and the design problem may be viewed now as
iFrom the numerical point of view (9.5) may be considered as a multi-dimension
version of the best regression selection problem. Stepwise regression and best
subset selection are the popular algorithms and can be easily adopted to solve
(9.5). In fact the same methods may be used when the matrix f i(Sp~,B) is
replaced by its true value; see comment in the conclusion of Section 8. Let
i.e. we want to minimize the variance of prediction at point z1. In this case
and it is a very standard problem of selection of N predictors from Nx - 1
candidates and there exist a numerous number of the statistical packages which
can be used to do that. The author is not familiar with multi-dimension versions
of the corresponding software products, which are needed for more complicated
criteria. The idea to use the least squares technique for selection of the most
informative subset of sensors was probably initiated by Megreditchan (1979).
.
DESIGN OF SPATIAL EXPERIMENTS xlv
The search for an optimal design &,J may be viewed in this setting as a parti-
tioning of X into two sets of given size NX - N and N . The latter must contain
the most information about the whole set X; see, for instance, Shewry and Wynn
(1987), who proposed using
where the subscript “p” indicates that the matrix contains only elements corre-
sponding to the points (sites) with no observations. When (9.7) is replaced by
its empirical version
then the following simple and intuitively attractive exchange-type procedure may
be used to construct &; see Fedorov and Hack1 (1994):
(a) Given ( N ~ = { x i s } ; find
k N
i+ = arg maxmin
Add the point x;+ to the design: [ ( N + I ) ~ = [ N s + xi+. (b) Find
where xz E J ( N + ~ ) ~ and delete the point 2;- from the design, i.e. construct
t N ( s + l ) = <(N+l)s - 3 i - e Retun to (a).
Briefly, the exchange procedure (a), (b) may be spelled out in the following
way: add to the design the worst explained sites and delete from it the best
explained sites. Apparently, the approach may be called “model free”: only
existence of first two moments of observed Y is assumed. That may attract many
practitioners. However in the search for an optimal network we are confined to
sites where the measurements have been previously made. In other words, the
xlvi
selection of the most informative subsets of sites (sensors, observing stations)
may be discussed, but we cannot consider the problem of optimal extension.
Estimation of a parameterized covariance. In many practical cases the design
region X is a continuous set and the covariance function has to be known every-
where at X . The most popular approach is based on the assumption that this
function is homogeneous and isotropic, i.e.
V ( z , d) = V ( r ) , T 2 = (3 - .E')T (z - z')
with the subsequent parsimonious approximation of function V ( r ) ; see e.g. Cressie
(1991), Marshall and Mardia (1985), M a t h (1986), Ying (1995). The approach
is frequently used in geostatistics, where a single realization of a random field is
available, and in particular in the "kriging method" paradigm.
Methods from in Sections 5-8 are essentially based on approximation of the
observed random fields by regression models with random coefficients. When
prior to design of a network there exist some historical observations, then one
may use the technique, which was developed for these models.It is expedient to
note that accurate knowledge of A or A, is useful but it is not as crucial as the
knowledge of a covariance function in the Sacks-Ylvisaker approach. In fact, in
basic optimization problems (5.13) and (6.2) the objective functions depend upon
the sum M(<) + Mo, where Mo is defined by A. For instance, in the case of (6.2)
and therefore the role of A diminishes when either a2 --+ 0 or N + 00. Moreover,
the simple dependence upon A allows to construct numerically optimal designs
for different matrices A to learn about their sensitivity with respect to A.
In the simplest case, when the observational errors are negligible, the following
estimators may be used:
j=1
DESIGN OF SPATIAL EXPERIMENTS xlvii
where 8 f E", X j is the set of points with observations yj(z;) , and
for all j = 1,. . . , IC. It is assumed that functions f(z) are known and
Subsequently,
Actually, it is more convenient to use the matrix An directly than the function
p ( x , 2') in all numerical procedures discussed in Sections 5 , 6.
When the observational errors are comparable with the variations of 6, then
(9.9) must be replaced with more sophisticated estimators, which are computa-
tionally much more demanding and complicate. Details and references may be
found in Spjotvill (1977) and Fedorov et al. (1993).
10. Space and time
In most spatial experiments, after the sites are selected measurements are usually
taken on some regular schedule, for instance, several times a day, or they are
continuously recorded. Generally, the response function may depend upon time.
Random errors can be correlated both in time and space. We consider only
the simplest case, where there is no spatial correlation, following the ideas from
Section 2. The generalization for more general models considered in Sections 3-6
is straightforward.
To adopt (2.1) for the time dependent response we assume that
xlviii
and
When p j j t = 6 j j t , then the information matrix of observations made accord-
ingly to the time schedule ( (d t l z ) may be presented in the following form
For measurements which are correlated in time,
where, for the sake of simplicity, we assume that supp((dt/z) is a discrete set
tl, t 2 ... 7 t77 and R(z ) = p j j l ( z ) i -
When the measure ( ( d t l z ) is fixed for each given z, then all the results from
Section 2 may be used, with obvious replacement the function $(z, t), which in
the standard case has the form
for all criteria satisfying assumptions (a) - (e), by the function
G(2 , t ) = @(e> - trm(z)A(S)*
For instance, for the D-criterion the sensitivity function m - f T ( z ) M - ' ( [ ) f ( z )
must be replaced by m - trm(z)M-'([). More details may be found in Atkinson
and Fedorov (1988), Fedorov and Nachtsheim (1995), and Spruil and Studden
(1979). Formally the time dependent observation may be treated as a vector-
observation case (see, for instance, Fedorov (1972), Ch. 5).
Evidently, introducing the time variable does not change the basic theory,
but makes all techniques, including computing of optimal designs, more time and
effort consuming. However there exist models and optimality criteria for which
DESIGN OF SPATIAL EXPERIMENTS xlix
optimal designs are the same both for the static and for the time dependent
cases. For instance, the latter is true for models with uncorrelated observations
and with separable variables, when
or where
and the selected criterion satisfies assumptions ( a ) - ( e ) from Section 2; see Cook
and Thibodeau (1980), Hoe1 (1965), Huang and Hsu (1993), Schwabe (1994,
1995).
When time is included explicitly in model, then the concept of sensor alloca-
tion can be extended and "mobile" sensors may be introduced. In this case design
consists of trajectories x;(t) f X,O 5 t 5 2'. The topic is beyond the scope of
this survey. A reader can find the results and references in Chang (1979), Fedorov
and Nachtsheim (1995), Titterington (1980) and Zarrop (1979).
Acknowledgement
I am most grateful to my immediate colleagues D. Downing and M. Morris for
their very constructive and effective help in preparing this paper. I thank B.
Wheeler for his numerous and very useful comments and suggestions.
11. References
[l] Achieser, N. I. (1956). Theory of Approzimation, Frederick Ungar, New York.
[2] Atkinson, A. C. and V. V. Fedorov (1988). Optimum Design of Experiments.
In Kotz, S. and Johnson, N.I. Encyclopedia of Statistics, Supplemental Vol-
ume. Wiley, New York pp. 107-114.
1
[3] Atkinson, A.C. and A. N. Donev (1992). Optimum Experimental Resign.
Clarendon Press, Oxford.
[4] Benhenni, K. and S. Cambanis (1992). Sampling Designs for estimating In-
tegrals of Stochastic Process. Ann. Statist., 20, pp. 161.196.
[5] Bickel, P. J. and A. M. Herzberg (1979). Robustness of Design Against Au-
tocorrelation in Time, I, Ann. Statist. 2, 77-95.
[6] Bishoff, W. (1992). On Exact D-optimal Designs for Models with Correlated
Observations, Ann. Inst. Statist. Math. 44, 229-238.
[7] Bishoff, W. (1993). On D-optimal designs for linear models under correlated
observations with applications to a linear model with multiple response.
JSPI. 37, 69-80.
[SI Brimkulov, U., Krug G. and V. Savanov (1986). Design of Experiments for
Random Fields and Processes. Nauka, Moscow.
[9] Butkovskiy, A. G. (1982). Green's Functions and Transfer Functions Hand-
book, Wiley, New York.
[lo] Cambanis, S. (1985). Sampling Designs for Time Series. In Handbook of
Statistics. Vol. 5, North-Holland, New York, pp. 337-362.
[ll] Cambanis, S. and Y . Su (1993). Sampling Design for Estimation of a Random
Process. Stochastic Process Appl. 46, 47-89.
[12] Chang D. (1979). Design of Optimal Control for a Regression Problem. Ann.
Statist. '7, 1078-1085.
[13] Cook, R. D. and L. A. Thibodau (1980). Marginally Restricted D-optimal
Designs. JASA. 75, 366-371.
E141 Cook, R.D. and C. J. Nachtsheim (1980). A Comparison of Algorithms for
Constructing Exact D-optimal Designs. Technometrics, 24, 315-324.
DESIGN OF SPATIAL EXPERIMENTS li
[15] Cook, R.D. and V. V. Fedorov (1995). Constrained Optimization of Exper-
imental Design. Statistics. 26, 129-178.
[16] Cressie, N.A. (1991). Statistics for Spatial Data. Wiley, New York.
[17] Currin? C., Mitchell, T., Morris, M., and D. Ylvisaker (1991). Bayesian
Prediction of Deterministic Functions, With Applications to the Design and
Analysis of Computer Experiments. JASA. 31, 953-963.
[18] Davis, P. J. and Rabinowitz, Ph. (1984). Methods of Numerical Integration,
2nd Edition. Academic Press, New York.
[19] Eaton, J.L., Giovagnoli A., and P. Sebastiani (1994). TR 598. School of
Statistics, University of Minnesota.
[20] Ermakov, S.M. (Ed.) (1983). iVathematical Theory of Experimental Design,
Nauka, Moscow.
[21] Ermakov, S.M. (1989). Random Interpolation in the Theory of Experimental
Design, Computational Statistics and Data Analysis, 8, 75-80.
[22] Fedorov, V.V. (1972). Theory of Optimal Experiments. Academic Press, New
York.
[23] Fedorov, V.V. (1989). Optimum Design of Experiments with Bounded Den-
sity: Optimization Algorithms of the Exchange Type. JSPI. 24, 1-13.
1241 Fedorov, V.V. and A.B. Uspensky (1975). Numerical Aspects of Design and
Analysis of Experiments. Moscow State University Press, Moscow.
[25] Fedorov, V.V., Leonov, S.L. and S. A. Pitovranov (1988). Experimental
Design Technique in the Optimization of a Monitoring Network. In Fedorov
V.V. and Lauter M. (Eds). Model-Oriented Data AnaZysis. Springer-Verlag,
New York, pp. 165-175
[26] Fedorov, V.V. and W. Muller (1989). Comparison of Two Approaches. In
the Optimal Design of an Observation Network. Statistics. 19, 339-351.
lii
[27] Fedorov, V.V. and W. Muller (1989). Design of an Air-Pollution Monitoring
Network. An Application of Experimental Design Theory. Osterreichische
Zeitschrift fur Statistik und Informatik. l9, 5-17.
[28] Fedorov, V.V., Hackl, P. and W. Muller (1993), Estimation and Experimen-
tal Design for Second Kind Regression Models. Informatik, Biometrie und
Epidemiologie. 24, 134- 151.
[29] Fedorov, V.V. and P. Hackl (1994). Optimal Experimental Design: Spatial
Sampling. Calcutta Statistical Association Bulletin.
[30] Fedorov, V.V. and C. J. Nachtsheim (1995). Optimal Designs for Time - Dependent Responses. In Kitsos L. P. and W. G. Mueller (Eds). Advanced
in Model-Oriented Data Analysis. Springer-Verlag, New York. pp. 3-12. 44, 57-81.
[31] Fisher, R.A. (1947). The Design of Experiments, 4th Edition, Hafner-
Publishing Co. Inc. New York.
[32] Gribik, P.R., Kortanek, K.O. and J. R. Sweigart (1976). Design a Regional
Air Pollution Monitor Network: An Appraisal of a Regression Experimental
Approach. In “Proceedings of the Conference on Environmental ModeZing
and Simulation”. EPA 600/9-76-016. Research Triangle Park, North Car-
olina, pp.86-91.
[33] Guttorp, P., Le, N.D. Sampson, P.D. and J. V. Zidek (1993). Using En-
tropy in the Redesign of Environmental Monitoring Network. In Patil, G.P.
and Rao, C.R. (Eds). Multivariate Environmental Statistics. Elsevier Science
Publisher, New York, pp. 175-202.
[34] Hoel, P.G. (1965). Minimax designs in two dimensional regression. Ann.
Math. Stat., 29, 1134-1145.
I351 Huang, M.L. and M. C. Hsu, (1993). Marginally Restricted Linear-Optimal
Designs. JASA, 35, 251-266.
RESIGN OF SPATIAL EXPERIMENTS liii
[36] Kanwal R. P.(1971). Linear Integral Equations, Theory and Technique. Aca-
demic Press, New York.
[37] Karlin, S. and W. Studden (1966). Tchebyshefl Systems: with Applications
in AnaZysis and Statistics. Wiley, New York.
[38] Kiefer, J. (1959). Optimal Experimental Design. JRSS (B) , 21. 272-319.
[39] Kiefer, J. (1974). General Equivalence Theory for Optimum Designs (Ap-
proximate Theory). Ann. Statist.. 2, 849-879.
1401 Kunert, J. (1988). Optimal Designs for Correlation in the Plane. In Dodge,
Fedorov and Wynn (Eds). Optimal Design and Analysis of Experiments.
North-Holland, New York. pp. 123-131.
[41] Lindley, D. (1956). On a Measure of Information provided by an experiment.
Ann. Math. Stat. 27, 986-996.
[42] Marshall, R.J. and K. V. Mardia (1985). Minimum Norm Quadratic Es-
timation of Components of Spatial Covariance. Mathematical Geology. l7, 517-525.
[43] Martin, R.J. (1986). On the Design of Experiments under Spatial Correla-
tion. Biometrika. 73, 247-277.
[44] Matkrn, B. (1986). Spatial Variation, Lecture Notes in Statistics. Springer-
Verlag, Berlin.
[45] Megreditchan, G. (1979). L'optimization des reseaux d'observation des
champs meterologiques. La Meterologie. 6, 51-66.
[46] Megreditchan, G. (1989). Statistical Redundancy as a Criterion for Metero-
logical Network Optimization. 0 sterreichische Zeitschrift fGr Statistik und
Informatik. l.9, 18-29.
liv
[47] Micchelli, C.A. and G. Wahba (1981). Design problems for optimal surface
interpolation. In Ziegler Z. (Ed.). Approsimation Theory and Applications.
Academic Press New York. pp. 329-248.
[48] Mikhailov, G.A. and A. A. Zhigljavskii (1989). Uniform Optimization of
Weighted Estimates of the Monte Carlo Method. Soviet Math. DokZ. 38, 523-526.
1491 Mitchell, T.J. (1974). (a) An -4lgorithm for Construction of D-optimal Ex-
perimental Design. (b) Computer construction of D-optimal “First-Order”
Designs. Technometrics. 16, (a)203-210, (b)211-220.
[50] Morris M.D., Mitchell T.B. and D. Ylvisaker (1993). Bayesian Design and
Analysis of Computer Experiments: Use of Derivatives in Surface Prediction.
Technometrics. 35, 243-255.
[51] Miiller W. and A. Pizman (1995). Design Measures and Extended Informa-
tion Matrices for Optimal Designs when the Observations are Correlated,
TR No.47, Dept. of Statistics, Univ. of Economics Vienna.
[52] Muller - Gronbach Th. (1993). Optimal Designs for Approximating the Path
of Stochastic Process. Preprint Nr.. A-93-14. Freie Universitiit , Berlin.
1531 Nachtsheim, C.J. (1987). Tools for Computer-Aided Design of Experiments.
Journal of Quality Control. 19, 132-160.
[54] E t h e r , W. (1985). Egective Observation of Random Fields. Teubner Texte
Zur Mathematik-Band 72. Teubner Verlag, Leipzig.
[55] Nguen. N.K. and A. J. Miller (1992). A review of some exchange algorithms
for constructing discrete D-optimal designs. Comp. Statist. and Data Anal-
ysis. 14, 489-498.
1561 Oehlert, G.W. (1995). The Ability of Wet Deposition Network to Detect
Temporal Trends. Environmetrics. fi? 327-339.
DESIGN OF SPATIAL EXPERIMENTS Iv
[57] Oehlert, G.W. (1995). Shrinking a Wet Deposition Network. To appear:
Atmospheric Environment.
[58] A. P&zman (1986). Foundations of Optimum Experimental Design. Reidel,
Dordrecht.
[59] Pilz, T. (1991). Bayesian Estimation and Experimentd Design in Linear
Regression Models. Wiley, New York.
[60] Pukelsheim F. (1993). Optimal Design of Experiments. Wiley, New York.
[61] Rabinowitz, N. and D. M. Steinberg (1990). Optimal Configuration of a Seis-
mographic Network: A St atistical Approach. Bulletin of the Seismological
Society of America. 80, 187-196.
[62] Fho, C.R. (1973). Linear statistical Inference and its Applications, 2nd Edi-
tion. Wiley, New York.
[63] Ripley, B.D. (1981). Spatial Statistics, Wiley, New York.
[64] Sacks, J., Welch, W.J., Mitchell, T.J., and H.P. Wynn (1989). Design and
Analysis for Computer Experiments. Statistical Science. &, 409-423.
[65] Sacks, J. and D. Ylvisaker D. (1966). Designs for regression problems with
correlated errors. Ann. Math. Statist. 2, 66-89.
[66] Sacks, J. and D. Ylvisaker (1968). Designs for regression problems with cor-
related errors, many parameters. Ann. Math. Statist. 39, 49-69.
[67] Sacks, J. and D. Ylvisaker (1970). Designs for regression problems with cor-
related errors, 111. Ann. Math. Statist. 4l, 2057-2074.
[68] Sacks, J. and D. Ylvisaker (1970). Statistical designs and integral approx-
imation. Proc. 12tgh Bieu. Sem. Canad. Math. Cong. pp.115-136. Canad.
Math Cong., Montreal.
[69] SAS/QC Software: Design of Experiments Tools. (1995). SAS Institute, Inc.
lvi
[70] Seber G.A.F. (1977) Linear Regression Analysis, Wiley, New York.
[71] Seber G.A.F. (1984) MuZtivariate Observations, Wiley, New York.
[72] Shewry, M.C. and H. P. Wynn (1987). Maximum entropy sampling. JournaZ
of Applied Statistics. l4, pp.165-170.
[73] Schumacher, P. and J. V. Zidek (1993). Using Prior Information in Designing
Intervention Detection Experience. Ann. Statist. 2 l , 447-463.
1741 Schwabe, R. (1994). Optimal Designs for Additive Linear Models, Preprint
No. 1-5-94, Freie Universitat Berlin, Fachbereich Mathematik.
[75] Schwabe, R. (1995). Designing Experiments for Additive Nonlinear Mod-
els. In Kitsos, C.P. and W.G. Mueller, Advances in Model-Oriented Data
Analysis. Springer-Verlag, New York. pp. 77-85.
[76] Silvey, S.D. (1980). Optimal Design. Chapman and Hall, London, New York.
E771 Spjotvill, E. (1977). Random Coefficient Regression Models - A Review.
Statistics. 8, 69-93.
[78] Spruill, M.C. and W. J. Studden (1979). A Kiefer-Wolfowitz Theorem in
Stochastic Process Setting. Ann. Statist. 1. 1329-1332.
[79] Stroud, A.H. (1975). Approximate CalcuZation of MuZtipZe Integrals.
Prentice-Hall, Inc., Englewood Cliffs, New Jersey.
[80] Szego, G. (1959). OrtogonaZ PoEynomiaZs. Amer. Math. SOC., New York.
[81] Titterington, D.M. (1980). Aspects of Optimal Design in Dynamic Systems.
Technometrics. 22, 287-300.
[82] Wahba, G. (1971). On the Regression Design Problem of Sacks and
Ylvisaker, Ann. Mtah. Statist. 42, 1035-1053.
[83] Wahba, G . (1974). Regression Design for Some Equivalence Classes of Ker-
nels. Ann. Statist. 2, 925-934.
DESIGN OF SPATIAL EXPERIMENTS lvii
[84] Wheeler, B. (1994). ECRIP: Version 6.0 for Windows. ECHIP, Inc.,
Hockessin.
1851 Wynn, H. (1982). Optimum Submeasures with Applications to Finite Popu-
lation Sampling. In Statistical Decision Theory and Related Topics 111, Vol.
2. Academic Press, New York. pp. 485-495.
[86] Ying, 2. (1993). Maximum Likelihood Estimation of Parameters under a
Spatial Sampling Scheme. Ann. Statist. 21. 1567-1590.
[87] Ylvisaker, D. (1975). Design on Random Fields. In A Survey ofStatistica1
Design and Linear Models. North-Holland, Yew York.
[88] Ylvisaker, D. (1987). Prediction and Design. Ann. Statist 15, 1-18.
[89] Ylvisaker, D. (1988). Bayesian Interpolation Schemes. In Optimal Design and
Analysis of Experiments. Dodge V., Fedorov V. V., and Wynn, M. (Eds).
North-Holland, New York. pp. 169-278.
[go] Zarrop, M.B. (1979). Optimal Experimental Design for Dynamic System
Identification. Springer-Verlag, New York.
1911 Zhigljavsky, A.A. (1988). Optimal Designs for Estimating Several Integrals,
in Dodge, Y., Fedorov, V. V. and Wynn, H.P. (Eds). Optimal Design and
Analysis of Experiments. Nort h-Holland, New York.
11921 Zimmerman, D.L. and M.B. Zimmerman (1991). A comparison of Spatial
Semivariagram Estimators and Corresponding Ordinary Kriging Predictors.
Technometrics, 33, 77-91.
DESIGN OF SPATIAL EXPERIMENTS lix
INTERNAL DISTRIBUTION
1. D. J . Alexander 2. C. K. Bayne 3. W. R. Corwin 4. D. F. Craig 5. T. S. Darland 6. C. S. Daw 7. D. J. Downing
8-12. V. V. Fedorov 13. L. J. Gray 14. H. W. Hayden, Jr. 15. S. K. Iskander 16. P. Kanciruk
17-21. M. R. Leuze 22. D. E. McCabe 23. R. C. Mann 24. J. G. Merkle
25. G. R. Moline 26. M. D. Morris 27. R. K. Nanstad 28. C. E. Oliver 29. G. Ostrouchov 30. R. L. Schmoyer 31. R. F. Sincovec 32. P. T. Singley 33. D. A. Wolf 34. K-25 Applied Tech. Library 35. Y-12 Technical Library 36. Laboratory Records - RC
37-38. Laboratory Records Dept. 39. Central Research Library 40. M & C Records Office 41. ORNL Patent Office
EXTERNAL DISTRIBUTION
42. Prof. A. C. Atkinson, London School of Economics, London WC2A 2AE, United Kingdom
43. Dr. Richard Beckman, Statistics Group Al, Los Alamos National Laboratory, MS F600, Los Alamos, NM 87545
44. C. Bolton, Nuclear Electric, Berkeley Nuclear Laboratories, Berkeley, Gloucester- shire, GL139PB, United Kingdom
45. Prof. R. D. Cook, School of Statistics, University of Minnesota, 1994 Bufford Avenue, St. Paul, MN 55108-6042
46. Prof. D. F. Cox, Nuffield College, Oxford OX1 lNF , United Kingdom
47. Prof. Dennis Cox, Department of Statistics, Rice University, Houston, TX 77251- 1892
48. E. Eason, Modeling and Computing Services, Suite 105,6400 Lookout Road, Boul- der, CO 80301
49. Prof. Sherwood Ebey, Department of Mathematics, University of the South, Se- wanee, T N 37375
50. Prof. P. H a d , Department of Statistics, Vienna University of Economics, Augasse 2-6, A-1090 WIEN, Austria
51. Dr. David Hall, Statistics, Systems Department, Pacific Northwest Laboratory, P.O. Box 999, Richland, WA 99352
lx
52. Prof. J . A. Hartigan, Department of Statistics, Yale University, Box 2179 - Yale Station, New Haven, CT 06520
53. Dr. Dan Hitchcock, Office of Scientific Computing, ER-7, Applied Mathematical Sciences, Office of Energy Research, U. S. Department of Energy, Washington, DC 20585
54. Dr. Fred Howes, Office of Scientific Computing, ER-7, Applied Mathematical Sciences, Office of Energy Research, U. S. Department of. Energy, Washington, DC 20585
55. Prof. J. Stuart Hunter, 503 Lake Drive, Princeton, N.J. 08540
56. Prof. Mark Johnson, Department of Statistics, University of Central Flordia, Orlando, FL 32816-0370
57. Dr. A. M. Liebetrau, Computational Sciences Department, Battelle-Northwest, P. 0. Box 999, Richland, WA 99352
58. Dr. J. Lijengren, Pacific Northwest Laboratories, P. 0. Box 999, Richland, WA 99352
59. Dr. Michael McKay, Statistics Group Al, Los Alamos National Laboratory, MS F600, Los Alamos, NM 87545
60. Prof. Lisa Moore, Institute of Statistics and Decision Sciences, Duke University, Durham, NC 27706
61. Dr. David Nelson, Director of Scientific Computing, ER-7, Applied Mathematical Sciences, Office of Energy Research, U. S. Department of Energy, Washington, DC 20585
62. G. R. Odette, Department of Chemical and Nuclear Engineering, Ward Memorial Drive, University of California, Santa Barbara, A 93106
63. Dr. David J. Pack, CSR Institute, 8889 Bourgade AV, Lenaxa, KS 66219
64. Mr. Brent Pulsipher, Computational Sciences Department, Battelle Northwest,
65. Dr. Jerome Sacks, NISS, P. 0. Box 14162, Research Triangle Park, NC 27709- 4162
66. Prof. A. F. Smith, Department of Mathematics, University of Nottingham, Uni- versity Park, Nottingham NGi' 2RD, England
67. Dr. Alan Solomon, P. 0. Box 227, Omer 84965, Israel
68. Dr. Daniel L. Solomon, Department of Statistics, North Carolina State University, P. 0. Box 5457, Raleigh, NC 27650
69. Prof. S. M. Stigler, Department of Statistics, University of Chicaago, 5734 Uni- versity Avenue, Chicago, IL 60637-1546
70. A. Taboada, Mail Stop TO-E10, Division of Engineering, U.S. Nuclear Regulatory Commission, Washington, DC 20555
71. M. G. Vassilaros, MS T10-E10, Division of Engineering, U. S. Nuclear Regulatory Commission, Washington, DC 20555
72. Prof. William Welch, Department of Statistics and Acturial Science, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada
P. 0. BOX 999, K1-86, Richland, WA 99352
DESIGN OF SPATIAL EXPERIMENTS lxi
73. Prof. Kee Wong Weng, Biostatistics, UCLA School of Public Health, 10833 Le
74. Jeff Wu, Department of Statistics, Mason Hall, University of Michigan, Ann Arbor,
75. Prof. Y. C. Yao, Department of Statistics, Colorado State University, Fort Collins, CO 80523
76. Prof. Don Ylvisaker, Department of Mathematics, University of California, Los Angeles, CA 90024
77. Prof. J. V. Zidek, Department of Statistics, University of British Columbia, 2021, West Mall, Vancouver, BC V6T 122 Canada
Conte Ave., Los Angeles, CA 900241722
MI 48109
78. Offjce of Assistant Manager for Energy b e a r c h and Development, Department of Energy, Oak Ridge Operations Office, P. 0. Box 2001, Oak Ridge, T N 37831-8600
79-80. Office of Scientific and Technical Information, P. 0. Box 62, Oak Ridge, T N 37830